Strata update doubles local inference speed for quad-R9700 setups: 70 to 140-160 tok/s
poofph · reddit · 2026-11-03
A Redditor benchmarked the latest Strata update's Flash Next engine and reports a major speedup for multi-GPU local inference.
Real-world numbers (four R9700 GPUs + 256GB system RAM, running unsloth Q4KXL quant):
- Before update: 70 tok/s output
- After update: 140-160 tok/s output, 3200 tok/s prefill
- Q5 quant fell 0.1GB short of VRAM for split engine mode, so it ran "normal" at only 70 tok/s
For reference, Radiance still leads on Flash Next at 235 tok/s output and 7500 prefill, but Strata is closing the gap and giving users more choices of models they can run locally.
More from Infra
- Hugging Face Kernels primer: load optimized CUDA kernels from the Hub in one line — ariG23498 · 2026-11-03
- Hash partitioning cuts Neo4j parallel import from 250.6s to 186.5s on 3.2M rows — JeremyCMorgan · 2026-11-03
- ai& x AMD host one-day Tokyo hackathon with ¥10.2M in credits, 750K JPY tokens per winner — DavidBennett__ · 2026-11-03
- Musk mocks compute-satellite heat dissipation skeptics: 'They'll never get a job at SpaceX' — elonmusk · 2026-11-03
- tinygrad's exabox drops to $5M: 256x MI300X, 669 PFLOPS, delivery Q3 2027 — andrew_n_carr · 2026-11-03
- Local GLM 5.3 Flash finishes the same coding task 4.2x faster than cloud (10:24 vs 43:48) — hyudryu · 2026-11-03